In the previous chapter, we saw how it's possible to use a fully connected neural network to approximate a nonlinear function. These types of networks suffer from one major problem: they have too many parameters to learn. This will not only increase the computational time, but also the chance of overfitting the data. Overfitting occurs when our model is not able to generalize outside the training data, and results in poor performance on new inputs. This is quite dangerous, because you might realize you are overfitting only after implementing the model in production.
There are many different neural network architectures that can counter this issue. The most common one, especially in computer vision, is the Convolutional Neural Network (CNN).
The following topics will be covered in this chapter:
- Understanding CNNs
- Convolutional...
